In movies there is always an image that can be zoomed into and enhanced to reveal car plate numbers for example or other details that are not really there when you zoom in beyond the native resolution of the image.
Details that are not in the image cannot be re-created but there is a finite number of probable scenarios that could result in the given image when scaled down. This approach will create several and sometimes many alternatives but with the addition of a pre-created sample database, a statistical engine and a set of configurable properties, the number of options can be manageable.
For example you have an image that has a blurred license plate number. You don’t have enough information to deduce the number from the given information but if you define this as a number plate blur and set some lightening and other settings, the enhancement process could investigate which number plates under the defined settings would create a similar blur.
There will be as I said before more than one solution but with the aid of a registration database the likely result list can be very small.
This can be done with many types of data items, each with its own set of statistics database, for example:
- License Plate - find number
- Face - find some ethnic classification for example
- Handwriting - recognize some words / numbers
- Signs - recognize sign by logo, label or color scheme
I thought about some of the algorithms needed for this a few years ago after I finished a project where I was involved with creating the blurred or anti-aliased version of text in various languages to enhance the way it looks of a wide range of devices. I thought about what it would take to bring it back to the original vector glyph.
I you want to develop and use / sell this kind of system you can contact me and I will help you get started.
